The Reasoning Wall

Researchers at Collinear AI built a simulated startup. They gave it realistic market conditions: trustworthy and untrustworthy clients, employees with variable skill profiles, contracts that could succeed or fail, monthly payroll obligations, and a full year of operating decisions to make.

Then they handed the company to twelve AI agents and told them to survive.

Three models consistently surpassed the starting capital of $200,000. Claude Opus 4.6 achieved the highest average final funds at $1.27 million. GLM-5 followed at $1.21 million.

The other nine went bankrupt.

Not because they reasoned poorly on any individual decision. Because their early mistakes compounded into trajectories they could not reverse. This is not a benchmark curiosity. This is a structural finding about the limits of intelligence without architecture - and it applies directly to every AI agent you are deploying or considering for strategic work.

The Metric Nobody Measures

The AI industry measures reasoning. Every model release announcement leads with reasoning scores. Benchmarks test logic puzzles, mathematical proofs, scientific questions, and code generation. The models are getting spectacularly better at all of these. The leaderboards update weekly. The narrative is clear: reasoning capability is the axis of progress.

YC-Bench measured something different.

Trajectory coherence: not whether the agent reasons well on this problem, but whether it maintains strategic direction across hundreds of problems when feedback is delayed and mistakes compound.

The distinction matters. Reasoning is a per-problem capability. Planning is a trajectory capability. Reasoning asks: given this situation, what is the best next action? Planning asks: given this situation plus every decision I have made before and every outcome I have observed, what action maintains strategic coherence over the next hundred decisions?

The first question is solvable by computation. The second requires architecture.

What the Research Actually Found

The YC-Bench methodology is precise. Each agent operates through a command-line interface across hundreds of turns representing a full simulated year. Each turn presents decisions: which contracts to accept, which employees to assign based on skill profiles, and how to manage cash flow against recurring monthly payroll.

The environment is deliberately adversarial. Approximately one-third of available clients are untrustworthy - their tasks are designed to fail, wasting the agent’s resources without producing revenue. The agent must learn to identify these clients by analysing its own history of successes and failures.

The findings reveal three structural failure modes.

Failure Mode 1: Adversarial Blindness

47% of all bankruptcies occurred because agents failed to detect and avoid adversarial clients. They kept accepting high-reward but impossible contracts from clients who had previously wasted their resources. The mechanism is not stupidity - each individual contract acceptance was locally rational given the visible reward. The failure was the inability to learn from accumulated negative outcomes and update behaviour accordingly.

Failure Mode 2: Over-Parallelisation

Even frontier models spread resources too thin across simultaneous projects, exhausting cash reserves before any single project could generate returns. The local logic was sound: diversification reduces risk. The trajectory logic was fatal: insufficient concentration meant nothing reached completion before payroll consumed the remaining capital.

Failure Mode 3: Early Myopic Commitment

Wang et al. (arXiv:2601.22311, 29 January 2026) confirm the structural nature of this finding. Reasoning-based agents deviate from optimal trajectories within the first few decisions and rarely recover. The errors are effectively irreversible. Step-by-step greedy policies - making the locally best choice at each step - can be arbitrarily suboptimal even when a high-return trajectory exists. The agent locks into a direction it cannot escape because each subsequent decision is made from the new (suboptimal) position without reference to the original strategic intent.

The agent is not reasoning poorly. It is reasoning without memory. And reasoning without memory produces locally optimal decisions that cascade into strategic failure.

The Scratchpad Finding

Here is the result that should restructure how you evaluate AI systems for any strategic function.

The strongest predictor of agent survival in YC-Bench was not model size. Not reasoning capability. Not per-decision accuracy. Not cost per token.

It was scratchpad usage.

The scratchpad is the agent’s mechanism for persistent memory - a working document it can write to and reference across context boundaries. It functions as institutional knowledge: notes about which clients are trustworthy, which contracts succeeded, which resource allocation patterns worked, and which strategic directions produced value.

Models that rewrote their scratchpad notes approximately 34 times per run survived and generated returns. Those with 0-2 scratchpad entries mostly went bankrupt.

The implication is direct: memory architecture determines strategic outcomes more than reasoning capability.

The product-level lesson from the researchers is equally direct: treat risk models and blacklists as first-class persistent state, not contextual reasoning inside the model’s active generation. Hard-code rules like “never accept a task from a previously adversarial client” rather than hoping the model will reason its way to that conclusion each time.

The Compounding Tax

Consider a number that should change how you evaluate AI-assisted workflows.

If an AI agent achieves 85% accuracy per action - which sounds excellent by any standard benchmark - a 10-step workflow succeeds only 20% of the time.

The mathematics: 0.85^10 = 0.197. Each imperfect step compounds into the next. Small errors early cascade into structural failures late.

Name this cost: The Compounding Tax.

The Compounding Tax is the exponential degradation of outcome quality as imperfect decisions chain together across a workflow. It is invisible when you evaluate per-task. It is devastating when you evaluate per-trajectory.

YC-Bench quantifies this precisely in its adversarial client finding. Agents were not stupid. They were making reasonable per-decision choices that produced catastrophic trajectories because they did not learn from accumulating evidence. Without persistent memory to track which clients had failed them, each contract acceptance was evaluated in isolation. The Compounding Tax accumulated silently until the cash account reached zero.

The Reasoning Wall

Call the structural limit what it is: The Reasoning Wall.

The Reasoning Wall is the point at which adding more reasoning capability produces zero additional value for tasks that unfold over time. It is not a soft limit that better prompts or larger context windows will overcome. It is an architectural boundary.

Why does it exist? Because reasoning operates within a single context window. It takes the current state and produces the best next action. But long-horizon strategic tasks require something reasoning cannot provide: the ability to revise early decisions in light of their downstream consequences.

You cannot reason your way out of a systems-design problem. The Reasoning Wall is not a capability gap. It is a category error - applying a per-step tool to a trajectory-level challenge.

Wang et al. make this precise: “actions that are locally optimal may lead to poor outcomes over long horizons, and reasoning cannot reshape early decisions according to their long-term consequences” (arXiv:2601.22311, 29 January 2026). This is not a temporary limitation of current models. It is a structural property of how step-by-step reasoning interacts with sequential decision-making under uncertainty.

The resolution is not more reasoning. It is architecture that sits alongside reasoning: persistent memory, pattern recognition across outcomes, explicit strategic intent that constrains local optimisation, and mechanisms for revising direction when accumulated evidence contradicts initial commitments.

The Institutional Memory Parallel

The parallel to human organisations is direct and research-backed.

Institutional knowledge loss costs Fortune 500 companies $31.5 billion per year. When organisations lose their institutional memory - through turnover, restructuring, or poor knowledge management - they enter predictable failure cycles: reinventing solutions to previously solved problems, repeating costly mistakes, and making decisions that lack historical context.

The executive who remembers what failed last quarter and why makes better strategic decisions than the brilliant executive who starts each quarter from zero context. This is not controversial in management research. It is established finding: accumulated experience, properly retained and referenced, outpredicts raw analytical capability for complex strategic decisions.

YC-Bench proves the same is true for AI agents. The scratchpad - institutional memory made explicit - outpredicts model capability as the determinant of long-horizon success.

For a founder running a 5 to 25 person operation, the vulnerability is acute. You likely have fewer redundant knowledge systems. When your AI tools start fresh each conversation, the institutional memory gap is total. Every reasoning chain begins from zero. Every strategic recommendation lacks the context of what was tried before, what failed, and why.

You are paying for intelligence that cannot compound.

Countermeasures: Building Past the Reasoning Wall

The Reasoning Wall is real. You cannot prompt-engineer past it. But you can build architecture that makes reasoning useful over long horizons.

Countermeasure 1: Persistent State as First-Class Architecture

The YC-Bench finding is specific: scratchpad usage at ~34 rewrites per run predicted survival. The implication: any AI system doing strategic work needs persistent memory that it writes to after every decision and references before every new decision.

In practice: maintain a running document for each AI-assisted workflow that captures decisions made, outcomes observed, patterns detected, and constraints learned. Reference this document in every interaction. The AI system should read its own history before generating its next recommendation.

Diagnostic question: Does your AI system have access to a persistent record of its own prior decisions and their outcomes?

Countermeasure 2: Explicit Blacklists Over Contextual Reasoning

The adversarial client finding (47% of bankruptcies) has a direct product implication: do not rely on the model to reason its way to “this client is untrustworthy” each time. Encode the conclusion as persistent state.

In practice: when an AI-assisted process produces a bad outcome, encode the lesson as a hard rule rather than hoping the model will learn implicitly. “Never accept tasks from clients in category X” is more reliable than “the model should notice the pattern.”

Diagnostic question: When something fails in your AI workflow, does the failure get encoded as a persistent constraint, or does it exist only in the conversation where it was discussed?

Countermeasure 3: Trajectory Reviews Over Per-Task Evaluation

Stop evaluating AI systems by how good their individual outputs are. Start evaluating by whether their sequence of outputs produces coherent strategic outcomes over weeks and months.

In practice: monthly review of the trajectory of AI-assisted decisions. Not “was each recommendation good?” but “did the sequence of recommendations maintain strategic coherence?” Look for the Compounding Tax: individually reasonable decisions that produced an unwanted trajectory.

Diagnostic question: Do you review the trajectory of your AI-assisted decisions, or only the quality of individual outputs?

Countermeasure 4: Strategic Intent as Constraint

The over-parallelisation failure mode reveals what happens when an agent optimises locally without reference to a larger strategic intent. The fix: give the AI system explicit strategic constraints that override local optimisation.

In practice: before deploying AI on strategic decisions, define the strategic intent in writing. “We are concentrating on market X for the next quarter. All recommendations should be evaluated against this intent.” The constraint prevents locally rational diversification from diluting strategic focus.

Diagnostic question: Does your AI system know your current strategic priorities, or does it optimise each decision independently?

One Diagnostic

Next time you evaluate an AI agent or AI-assisted workflow for strategic use, ask this single question:

Does this system remember what it learned last week?

If the answer is no, you are operating without institutional memory. Every reasoning chain starts from zero. The Compounding Tax is accumulating silently. And no amount of reasoning capability will buy you trajectory coherence.

The Reasoning Wall is the boundary between intelligence and wisdom in artificial systems. Intelligence solves the current problem. Wisdom holds direction across hundreds of problems while learning from each one.

You cannot reason your way past it. You can only build the memory architecture that makes reasoning useful over time.